Voice Agents for Insurance GTM

This is a working transcript with chapter timestamps. It may contain transcription errors. Private names have been removed; check the video before quoting a precise phrase.

00:00:00 The insurance data problem

00:00:00 Hey everyone, my name is Karim. I'm a team engineer at Adapt, and we automate the busy work at insurance. I'm QR Code for you guys, I don't think anyone here is the first one to know, unfortunately. When I saw this submission statement when I was interviewing for Adapt, and I didn't really connect with it, I don't know, maybe because I'm Australian, and when I go to the doctor for a visit, I don't have to bring my insurance card, so insurance is not really something I think about too much when I'm working at Adapt.

00:00:33 But interestingly, when I was interviewing for Adapt, I was connected with one of the senior AEs of the team, and what the first time I heard was like, if there was one piece of information or intelligence on the accounts that are relevant to you guys with your ICP accounts, what would you want? What's something that's really difficult to get? And he told me a few things. He was like, oh, the calories that the insurances work with, or AMS, which is my CRM template of the insurance world. Also the policy count, right, the number of policies they

00:01:04 manage. And I'm thinking to myself, like, you know, as a, you know, jigsaw engineer that works for companies of different sizes, it's like, all the time, it's like these, you know, tech staff related pieces of information. They're pretty easy to get. You just scrape job posts, you know, you look at their website, blog articles, LinkedIn, you know, so I thought, you know, well, like, this seems pretty easy, and so I joined Adapt. First thing I did, as anyone would, is went to the CRM, and I wanted to make sure, like, our accounts were, you

00:01:36 know, had good quality, right? Like, I think when it comes to, like, jigsaw engineering, or robots in general, is making sure that, you know, your TAM, your ICP accounts have really good hygiene, really good account data, because that's kind of like downstream of everything, right, like signaling, targeting, segmentation, etc. And so, a lot of the accounts look like this, okay? But fortunately, I'm like, okay, I've been in this situation before,

00:02:05 Researching account websites

00:02:06 and, you know, I've got that domain there, so just go online to the domain, push it through to, like, you know, some sort of domain resolution endpoint, there's so many of them, you just have to use a Deepline play, like, this is kind of like a walk in the park. And so, like, the first thing I was thinking also, too, was like, okay, I'm going to go to LinkedIn, because I have to use that to, like, scrape and get some more insights, right? So I went to LinkedIn, and, yeah, so there was no LinkedIn account. That was weird, I found, like,

00:02:36 maybe that's just, like, one example, fair enough. I'm going to go to the domain, right? Domain is reliable, that's, like, a useful kind of information. I'm pretty sure this was created before I was born. So, yeah, I realized, like, a lot of the things, when it came to, like, I can pull up on, like, LinkedIn's deep learning engineering, and I just kind of throw out the window, and think in first principles, and, like, even before getting to that point of, like, I have to rethink this, I bought a different technique,

00:03:06 so I was like, okay, what I'm going to do is, like, scrape all the domains, every single page, I'll use context there, and context there is similar to, like, Bible, it makes it really easy and cheap to just scrape web pages, and I can just create a Deepline play, and just scrape every page, send that to, you know, Brandon, I completely agree with you, it's very cost effective, and so I just kind of went for that, and a lot of the websites look like this, and there wasn't much information, and so I really had to rethink how I can approach this,

00:03:36 and it made me realize I had to go beyond the public web, and potentially explore voice agents, okay? So, just a show of hands, who here has played around with voice agents? Okay, keep your hand up if you think they could be useful for GTM, okay? And keep your hand up if you are currently using the introduction today. No? Oh, okay. Like, a few people, okay? So, my thinking coming into this was like, okay, you know, voice agents are really interesting,

00:04:14 because it's not like this only exists in the world, it does exist in the world, right? It either exists in the minds of people, like, in terms of the EMS, voice account, obviously information that's relevant to us, or, like, internal databases,

00:04:25 How the voice agent works

00:04:27 and so it's like, how can we extract this? And it's like, okay, we're talking to people, right? We have CISA, right? We don't want them spending time to, like, figure out if they're ICP fear over the phone, right? That's what was happening before. It was like, we're spending several hours per day speaking to prospects that were definitely not ICP fear, like, they weren't using AMS that we supported, or they weren't preparing us, that we didn't enter it with, et cetera, et cetera. And so I thought voice agents would be interesting, but I also thought, you know,

00:04:55 this is something that's, like, six months away. It's only six months away. And the problem was I didn't have six months, right? And so, you know, they usually say, well, like, constraints breed creativity, and so I'm like, all right, I'm going to get my head down and figure out how I can leverage voice agents. So just to be, like, one-on-one, so we're all on the same page here, I want to talk a bit about voice agents. So there's, like, two primary ways to use voice agents.

00:05:19 I like that this is sort of the architecture behind voice agents. So you have the cascaded workflow, and you have speech-to-speech. Cascaded is what is, like, most common. If you've ever, like, actually spoken to a voice agent on the phone, it's likely it's the cascaded flow. And basically it's pretty simple. It's person talks on the other end. That audio gets transcribed by something we call an ASR model, also speech-to-text is another term for it, STT. So speech gets transcribed into text, and that text gets sent to an LLM with a given system instruction.

00:05:50 And then that output, the LLM, gets sent to a text-to-speech model. This is where ElevenLabs comes in, right? ElevenLabs has, you know, very popular TTS models. There's many others in the space. And then that speech gets sent to the person that keeps running the loop. Okay? Now, there's a lot of benefits with this sort of model, but there's also some downsides. Benefits is it's very configurable, right? So there is a lot of different ASR models, and you can run it through an eval, so depending on, like,

00:06:17 if the industry works with the terms being used, the word error rates with WDR, so like the sort of the benchmark, like the support metric with ASR models, that, you know, depending on, like, sort of how those models perform, you can kind of switch in, switch out. Same goes with LLM. Same goes with TTS. So they're very configurable. Another benefit is there's just a lot more guide rails and configurability. Okay. Yeah, there's a lot more, like, guide rails that you can kind of put into the LLM.

00:06:46 So, like, once the LLM sends a response, you can have somebody that's kind of monitoring those responses, and if you don't like it, you can kind of, like, stop it, right? And so that's really useful, and this is probably a big reason why it probably sits more, like, in healthcare and financial services where it's more mission critical. The downside is there's more latency. It has to go through all these different, you know, providers to get to, you know, to get to them and stop it. And then also, here's the big problem,

00:07:10 is that you lose a lot of context. Everything gets transcribed into text, so you lose the emotion, the tone, et cetera. Speech is a bit different. So it's, like, I talk, so the prospect talks, so the person on the other end talks, and that is received. Those native speech tokens are received, and the model outputs throughout speech. So this means that you're maintaining the voice, the tone. It also means latency is better, and it also means that, you know, so the downside is that there is, it's hard to be configurable.

00:07:42 You can't do a whole, you know, monitor the LN query, right? It's more of a black box. So this is the pros and cons. At the time, about two months ago, we chose to go with the standard because it was more configurable, and the speech models that we had at the time were just still, it was too nascent of a technology, okay? So let me show you an example of some of those polls. I'll actually give you more context.

00:08:05 Live call demo

00:08:05 So the first thing I wanted to do, again, was, like, I wanted to survey our prospects. I also wanted to call business lines because that's compliant. You know, call up the business lines and just ask them questions and see if, like, maybe they're willing to give out the information. Maybe not. Maybe it's all a waste of effort. So let me show you one of the calls. So this is one of the calls.

00:08:52 What are the main barriers you guys work with? Okay, well, I'm realizing that I don't have really much time. Okay, so I'm going to power through this. Okay, so that one worked really well. But then we're like, okay, top of funnel is, like, you know, not too strong, so we want to get more leads in. So I'm like, okay, let's look back at the speech models. And I realize, oh, that was a fun one. Okay, so I realize, so opening an eye, the list is GPT, real-time.

00:09:16 One model was speech-to-speech model. It's full-duplex meaning it thinks while it talks. And there's two things that are really important when it comes to an app by a voice agent that's doing prospecting. One, you want to make sure you understand the emotion, the context behind it. Actually, coming back to this is really important. If someone says a prospect after you pitch them, they're like, hmm, sure, versus hmm, sure. Those are, like, two different responses, right? But that gets transcribed the same way when it comes to the standard model,

00:09:43 right? And so also each option handling is really about the standard models. Anyway, that was the best approach for that. I'm going to show you, because we booked nine meetings in the first two days of deploying this new voice model, I'll show you a little bit of it. So the background audio is like... Excuse me, it's irrelevant. Thanks. So the reason I'm calling is Adapt helps agencies deal with carrier notices and documents that don't come through normal downloads.

00:10:28 Carrier notices for cancellation, do you mean? Yeah, cancellation, reinstatements, things like that. Often they land in emails. And then like the objection handle is here. It's an objection handle. Okay, yeah, that makes sense. Yeah, so it ends with an additional, and then towards the end right here, it's funny.

00:11:04 I've been booked in five and a half minutes. Yeah, that's it. A lot of fun stuff. Here's some stats. 70-plus accounts were breached from like 25% through a survey agent. Ten meetings booked in the first big launch. $80 per meeting booked. And that keeps dropping down as we improve latency, improves the bang of the objection handle.

00:11:35 Questions

00:11:38 A lot of fun stuff. That's it from me. I'm over time. Thanks. I'm not a cop. I knew this question would come up. But I think it's illegal to AI a robot cop. So yeah, okay, so here's the thing, I'm not a lawyer. I don't work in insurance, I'm not a lawyer either. But my understanding is that it's legal to call up the business line. And we only call up the business line.

00:12:13 Was that speech to speech? Yeah, when I just showed that to you in real time, one. Also shout out to [provider name unverified]. They're doing all the voice info, so like telephony, numbers. They're kind of like, you can think of them as like the Deepline for voice agents. Really recommend it. [Provider name unverified] is not AI. Shout out to them.